Inside an s6 container, `gateway run` redirects to the supervised
gateway and then keeps the CMD process alive as a no-op heartbeat so
/init doesn't start stage-3 shutdown. That heartbeat is
`os.execvp("sleep", ["sleep", "infinity"])`, which does a PATH lookup
for the `sleep` binary. When PATH was empty/truncated/clobbered at that
point — e.g. after user customizations rewrote PATH, or on a minimal
image without `sleep` on PATH — the exec raised FileNotFoundError,
killing the CMD process and causing /init to tear down every service:
the container failed to start (issue #36208, a regression in the s6
image from 2026.5.28).
Wrap the exec in try/except OSError: on success it still replaces the
process with the cheap `sleep` heartbeat (no resident Python
interpreter, and the existing process-tree/recursion contract is
preserved); on failure it falls back to `_block_until_terminated()` —
a SIGTERM handler (clean 128+signum exit on `docker stop`) plus a
signal.pause() loop, which needs no external binary and so can't fail
on PATH state. A threading.Event().wait() fallback covers platforms
without signal.pause().
Keeping execvp as the primary path (rather than replacing it outright)
preserves the `sleep infinity` heartbeat that the docker integration
tests assert (test_gateway_run_supervised.py) and avoids leaving a
full Python interpreter resident for the container's lifetime.
Verified end-to-end on a built image: with execvp forced to fail,
_block_until_terminated() blocks cleanly instead of raising
FileNotFoundError; normal boot still runs the cheap `sleep infinity`
heartbeat; the 6 test_gateway_run_supervised.py integration tests pass.
Salvages the two community fixes for this issue — the fallback design
from #36221 (@Pluviobyte) and the signal.pause() heartbeat from #36267
(@karmeleon) — and adds regression tests for both the normal and
sleep-missing paths.
Co-authored-by: Pluviobyte <Pluviobyte@users.noreply.github.com>
Co-authored-by: karmeleon <karmeleon@users.noreply.github.com>
Closes#36208.
eee32cdd52